Implicit Neural Radiance Field (NeRF) techniques have been widely applied and shown promising results for scene decomposition learning and rendering. Existing methods typically require encoding spatial and semantic coordinates separately, followed by deep neural networks (MLP) to obtain representations of the entire scene and individual objects respectively. However, these implicit neural field methods mix scene data and differentiable rendering together, which results in issues with expensive computation, low interpretability and limited scalability. In this article, we propose NeRF-IS (Explicit Neural Radiance Fields in Semantic Space), a novel 4D neural radiance field model architecture, that integrates 3D space and semantic space modeling, which can perform both scene-level and object-level modeling. Specifically, we design a hybrid method of explicit spatial modeling and implicit feature representation, which enhances the model’s ability in scene semantic editing and realistic rendering. For efficient training of NeRF-IS, we apply low rank tensor decomposition to compress the model and speed up the training. We also introduce an importance sampling algorithm that uses a volume density prediction network to provide more accurate samples for the whole system with a coarse-to-fine strategy. Extensive experiments demonstrate that our system not only achieves competitive performance for scene-level representation and rendering of static scene, but also enables object-level rendering and editing.

Differentially Private Stochastic Gradient Descent (DP-SGD) is a prime method for training machine learning models with rigorous privacy guarantees. Since its birth, DP-SGD has gained popularity and has been widely adopted in both academic and industrial research. One well-known challenge when using DP-SGD is how to improve utility while maintaining privacy. To this end, recently we have seen several proposals that clip the gradients with adaptive thresholds rather than a fixed one. Although each proposal comes with some theoretical justification, the theories often rely on strong assumptions and are not compatible with each other. It is hard to know whether they are good in practice and how good they are. In this paper, we investigate adaptive clipping in DP-SGD from an empirical perspective. With extensive experiments, we were able to gain some fresh insights and proposed two new adaptive clipping strategies based on them. We cross-compared the existing methods and our new strategies experimentally. Results showed that our strategies did provide a substantial improvement in model accuracy, and outperformed the state-of-the-art adaptive clipping methods consistently.

While gradient aggregation playing a vital role in federated or collaborative learning, recent studies have revealed that gradient aggregation may suffer from some attacks, such as gradient inversion, where the private training data can be recovered from the shared gradients. However, the performance of the existing attack methods is limited because they usually require prior knowledge in Batch Normalization and could only reconstruct a single image or a small batch one. To make the attacks less restrictive and more applicable, we propose an effective and practical gradient inversion method in this paper. Specifically, we use cosine similarity to measure the difference of gradients between the synthesized and ground-truth images, and then construct an input regularization for the fully connected layer to ensure the fidelity of the image. Moreover, we apply the total variation denoising strategy to the convolution feature map for further improving the smoothness of the reconstructed image. Experimental results demonstrate that our method can reconstruct high fidelity training data on a large batch size for complex data sets, such as ImageNet.

Graph Convolutional Networks (GCNs) are powerful representation learning methods for non-Euclidean data. Compared with the Euclidean data, labeling the non-Euclidean data is more expensive. Meanwhile, most existing GCNs only utilize few labeled data but ignore most of the unlabeled data. To address this issue, we design a novel end-to-end Iterative Feature Clustering Graph Convolutional Networks (IFC-GCN) that enhances the standard GCN with an Iterative Feature Clustering (IFC) module. The proposed IFC module constrains node features iteratively based on the predicted pseudo labels and feature clustering. Further, we design an EM-like framework for IFC-GCN training, which improves the network performance by rectifying the pseudo labels and the node features alternately. Theoretical analysis and experimental results show that our proposed IFC module can effectively modify the node features. Experimental results on public datasets demonstrate that IFC-GCN outperforms state-of-the-art methods on the semi-supervised node classification task.